A Diffusion Model based User Embedding Generation for Cold Users 


Vol. 14,  No. 7, pp. 562-568, Jul.  2025
https://doi.org/10.3745/TKIPS.2025.14.7.562


PDF
  Abstract

The cold-start user problem refers to the degradation in recommendation performance caused by insufficient user-item interaction data in collaborative filtering-based recommender systems. As an initial study that addressing this issue using the diffusion models, we propose a novel approach that leverages the diffusion model’s denoising capability to enhance the quality of embeddings for cold-start users. Specifically, we utilize a pre-trained collaborative filter to compute the average embedding of items that a user has interacted with and use it as a condition for the diffusion model. The model then generates a refined user embedding aligned with the collaborative filter’s learned space, which is subsequently used to compute preference scores via inner products with item embeddings. Experimental results on two real-world datasets show that our method outperforms comparison methods in item recommendation for cold users with fewer than 40 interactions, achieving higher Precision@20, Recall@20, and NDCG@20 metrics.

  Statistics
  Cite this article

[IEEE Style]

J. Han and H. Lee, "A Diffusion Model based User Embedding Generation for Cold Users," The Transactions of the Korea Information Processing Society, vol. 14, no. 7, pp. 562-568, 2025. DOI: https://doi.org/10.3745/TKIPS.2025.14.7.562.

[ACM Style]

Jungkyu Han and Hak-Sung Lee. 2025. A Diffusion Model based User Embedding Generation for Cold Users. The Transactions of the Korea Information Processing Society, 14, 7, (2025), 562-568. DOI: https://doi.org/10.3745/TKIPS.2025.14.7.562.

Forms

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)

Old Journals

Indexing

All publications of TKIPS is indexed in DOI, EBSCO, Google Scholar, Crossref, and CrossCheck.

TKIPS is also selected as the Journal for Accreditation by NRF (National Research Foundation of Korea).

Related Journals